Automatic pulmonary organ and small nodule segmentation in CT scans: basing on k-means and DU-Net++

Abstract

Accurate pulmonary nodule segmentation in Computer Tomography (CT) scans is significant in clinical treatment of lung cancer. In this paper, an approach based on k-means clustering is proposed for lung organ segmentation in order to remove irrelevant chest tissues. Then a convolution neural network (CNN) model, Dense U-Net++ (DU-Net++) is constructed for detecting and segmenting nodules. The model includes three parts: down-sampling by DenseNet201 for feature extraction, up-sampling by trainable deconvolution for image restoration and middle layers by skip connection for feature fusion. The public dataset, LIDC-IDRI, is used for training and testing and the performance of DU-Net++ achieves 96.0% and 91.3% in Dice coefficient on the training set and validation set. The Dice coefficient on small nodules in validation set is 87.56%. The results indicate that the proposed model can offer a correct segmentation reference to doctors, reducing the time as well as the pressure of reading CT scans.